I really don't have much to say

Claude Design fundamentally altered my understanding of AI

“I’m glad the partnership is unlocking the part of the work that felt heaviest. Let’s get the widget from design B into your codebase…” - Claude Design.

That was Claude Design’s response when I said that working with it was “absolutely thrilling.” It was disconcertingly accurate in summarizing why I find it so thrilling - it unlocked (really, unblocked) the part of the work that felt heaviest to me.

As a long-time programmer, I have been using Claude Code for a while now, but that interaction has always been complicated. Because I can do the work myself, it neither unlocks nor unblocks it. I hope it makes it more efficient. But is it really more efficient in the long term? Am going to end up with a codebase I can understand and maintain? Do I enjoy my work less in the bargain? Those questions yield complex feelings.

But my work with Claude Design is fundamentally different. I cannot create beautiful designs. My designs are embarrassing and clunky. I can see that they are bad but I do not know how to make them better. Claude Design can look at my existing work, and make it vastly, immediately, obviously better.

Claude Design reminds me of the David Ogilvy quote, “If each of us hires people who are smaller than we are, we shall become a company of dwarfs. But if each of us hires people who are bigger than we are, we shall become a company of giants.”

With Claude Design, relative to myself, I have unequivocally hired a design giant - for $200 a month - and it is absolutely thrilling.

My hobby app has been blocked on embarrassing design

For several years now, I have been working on a recipe management and shopping app that my wife and I use daily. Although I don’t think the world really needs yet another recipe app, I think this particular app might be useful to a very specific subset of people. For some time, I have had a specific idea for how to market the app to exactly those people.

But I looked at the app’s design, and I found it embarrassing. It was clunky. It looked like a programmer created it. So, I haven’t marketed it.

What were my past solutions to the design problem?

The design problem is something I have had in the past with small projects. I have typically solved it in two ways - using a framework, and using 99 Designs.

Framework: Flutter/Material Design

I choose Flutter in part because it was cross platform, but also because it had a large library of Material Design widgets. As embarrassing as my app’s design is, it would have been far worse if I was creating widgets from scratch.

Material design gave me a lot of good ideas. But it also limited me to a sort of ‘paint by numbers’ approach. And although its basic typography and color schemes were supposed to help me create something sophisticated, that didn’t happen in practice.

99 Designs

When I have used 99 designs for other projects, I haven’t had bad results with it, but there were disadvantages.

How I use Claude Design to Prototype and then Implement Flutter Designs

Currently, I have been working with Claude Design by taking screenshots of my app in the simulator, and uploading them. I tell it I am happy with the functionality in the app and want to keep it constant, I just want to make it look better. I initially expressed my concerns about the app seeming clunky and unattractive.

Claude then produces several mockups of improved versions of each screen. It also produces an explanation of what it is “trying to do” with the design in the chat, and how its various mockups differ from a design perspective. It produces new prototypes in less than thirty seconds.

Its designs, straight out of the box, are clearly better. They often don’t quite work from a functional perspective, but they are almost always easily adapted into something that does.

When I see a design for an element or area of the screen that I like, I ask Claude Design to generate a Flutter/Dart snippet for that particular element. I then incorporate that element into my code. Usually the UI elements correspond to existing subroutines or widgets in the code, so this workflow is pretty efficient.

Claude Design is very good at generating Flutter code from its native JSX mockups. In the rare cases where it is slightly off, I then switch to Claude Code, explain to it what is wrong (e.g. the divider at this line of code should be bottom aligned, not top aligned), and it has always fixed it immediately.

I have now updated several complex screens this way, and those screens are vastly improved. The screens go from embarrassing to “that looks like a professionally designed app.”

There is some serendipity too

Although that workflow by itself is transformative for me, there have also been moments of serendipity, where it feels like I am collaborating more deeply with Claude Design.

It sometimes creates UI elements which don’t work, but when I see them I realize that something similar would work. I go back and forth with it, explaining how we can refine or alter the concept to make it viable. Because the turnaround is so fast, and the execution of my verbal ideas is so excellent, I feel like I can explore new ideas just to see if they will work out.

I heard once that you need to introduce some entropy into your queries to get more creativity from generative AI, but I find that Claude introduces entropy for me, sparking new ideas in my mind. That is when it feels like collaboration.

It is imperfect, and that doesn’t matter!

There are minor imperfections. For example, despite the color scheme of my app being blue/green, it periodically tries to insert terracotta as the primary color in the code snippets (not in the design, just randomly in the code). Here is some code it recently generated, where it refers to the “terracotta slot” in a comment, but the color is actually a dark blue, and it says that right in the comment.

static const _primary = Color(0xFF3F5A8A); // “terracotta” slot - deep indigo

In the chat associated with the code generation, it will literally be like “Where’s the terracotta, do we need to add it back to your theme?” I just tell it we are not using terracotta and we continue.

But those imperfections have helped me understand, in a way that I can’t with code, how much AI imperfections don’t matter when the result you are getting is so much better than your baseline.

Why I wrote this blog post: Thank you Claude Design Team

I have never written a blog post before, and I don’t intend to write any more. My hope is that maybe a human who works on Claude Design will see this and take it as a thank you.

If you know someone who works on Claude Design product, please forward this post to them.

I started using Claude Design the first day it came out. And it just worked. I was immediately productive. I started by trying to improve my sign in screen (because it was simple and low stakes), and I immediately realized - this could be the solution to my overall design problem. And it has been.

I probably have a few more weeks of work to improve everything in the existing app. Right now the interaction feels incredibly fertile, with one improvement opening space for another.

When I am done, I am finally going to market my app. Maybe it will find a small, niche market, or maybe it won’t, but I won’t be embarrassed to show it to my target customer. Thank you to the team that created Claude Design!

This Experience Changed my Perspective on Generative AI

This experience has transformed my understanding of generative AI in five key ways:

  1. Returning to David Ogilvy’s quote about hiring giants, I’ve always found it pithy, but misleading in the sense that it is obvious who is tall, but it is not obvious who can be a “giant” in your company. Additionally, even when you can identify “giants”, there is a lot of competition for them. Hiring good people is hard. But if the relative flexibility of the US labor market over Europe’s has created a significant difference in GDP growth over time, what will happen when everyone can hire amazing AI employees at zero risk, fire them at will, and find public feedback on which AI employees are the most productive? What happens when hiring giants is so easy? 


  2. I had read commentary that suggested coding was “uniquely amenable” to generative AI because it is verifiable (e.g. unit tests, etc). I saw a paper arguing that “verifiability” would be the deciding factor in where AI was applied effectively. But Claude Design is not verifiable. Its subjective and judgement based. And it works. It now feels to me as though the potential future for GenAI is much broader.


  3. I have read and heard so much about AI risks, and they are important. But I look at Claude Design, and I think “this is fabulously useful and has zero chance of killing us all.” How many additional tools can be like Claude Design - self-evidently not dangerous, just useful?


  4. I was previously worried about a bubble (and there may be speculative market crashes ahead, or specific companies that implode). Now, I feel that fundamentally, the next 10 years will be marked by many things getting suddenly better and more efficient. I think there will be great disruption in supply chains, and in the internal organization of companies, but it suddenly seems to me we are in for a turbocharged period of economic improvement - of actual things getting better rapidly.


  5. I now understand that when AI is augmenting your skills rather than replacing them, it is delightful and empowering to use. Suddenly an AI future seems a lot more fun to me. Perhaps its a future where some of your skills got less valuable, but now you have a much broader set of skills to apply to problems that interest you. For that reason, if you are a programmer, try it (and if you are a designer stressed by Claude Design, try Claude Code).

Addendum: Minor Requests for Improvements

If a Claude Design employee saw this, here are my very minor requests. To be clear, the tool is awesome and I would happily pay for it without any of these improvements.

  1. Periodically, it will tell me to save tokens by starting a new chat, but then I have to go back manually and try to establish relevant context. I wish there was a button to “compress important content into the start of a new chat.”

  2. Navigation in the prototype / artboard area:


Again, those are super minor suggestions, the tool as-is is incredibly productive for me.